Investigating potential future changes in surface water flooding hazard and impact

Surface water flooding (SWF) is a recurrent hazard that affects lives and livelihoods. Climate change is projected to change the frequency of extreme rainfall events that can lead to SWF. Increasingly, data from Regional Climate Models (RCMs) are being used to investigate the potential water‐related impacts of climate change; such assessments often focus on broad‐scale fluvial flooding and the use of coarse resolution (>12 km) RCMs. However, high‐resolution (<4 km) convection‐permitting RCMs are now becoming available that allow impact assessments of more localised SWF to be made.


| INTRODUCTION
Surface water flooding (SWF), also known as pluvial flooding, is a global hazard. For example, parts of Europe (Germany; Bung, Oertel, Schlenkhoff, & Schlurmann, 2011;and Italy;Di Salvo, Ciotoli, Pennica, & Cavinato, 2017) and Asia (Japan; Bhattarai, Yoshimura, Seto, Nakamura, & Oki, 2016;and India;Akilan, Balaji, Abdul Azeez, & Satyanarayanan, 2017) have all experienced SWF in recent years. SWF occurs when rainwater cannot drain away quickly enough through drainage systems or by soaking into the ground; instead, water lies on or flows over the ground. This type of flooding tends to occur as a consequence of intense and localised rainfall often associated with convective events (e.g. thunderstorms); however, it can be caused by prolonged 'moderate' rainfall (e.g. June 2007 flooding in Hull ;Falconer et al., 2009) sometimes with embedded high-intensity rainfall cells, or by rapid melting of snow.
Like all flooding, SWF causes significant disruption to people's lives and livelihoods, damaging homes and businesses (Bhattarai et al., 2016), closing roads (DfT, 2014), schools and hospitals, and disrupting water and power supplies and communications (Defra, 2018). It can also cause environmental and human health impacts and even deaths (Burton, Rabito, Danielson, & Takaro, 2016;Milojevic, 2015). Urban areas are particularly vulnerable due to the concentration of people, buildings, infrastructure, and associated impermeable surfaces, which results in overwhelmed drainage systems from increased surface runoff (Kaźmierczak & Cavan, 2011). Rapid urbanisation (UN, 2014) is likely to increase vulnerability in urban areas into the future (Houston et al., 2011;Miller & Hutchins, 2017).
SWF threat is also likely to increase due to climate change (Miller & Hutchins, 2017), with precipitation patterns predicted to shift towards more intense events in the future (IPCC, 2013). Until recently, climate models have been too coarse to assess the impacts of sub-daily rainfall that is a key driver of SWF (Miller & Hutchins, 2017); however, high-resolution convection-permitting regional climate models (RCMs) are now becoming available Pan et al., 2011;Prein et al., 2015) that allow impact assessments of more localised SWF to be made.
Traditionally, SWF forecasts are based on the probability of forecast rainfall exceeding a given threshold for a set of durations.
However, employing a hydrological model to estimate surface runoff has the potential to provide benefits beyond existing rainfall depth threshold approaches because the hydrological model enables the dependence on land cover, soil type, and antecedent soil moisture to be included (Cole, Moore, Wells, & Mattingley, 2016). Kaspersen, Ravn, Arnbjerg-Nielsen, Madsen, and Drews (2017) used runoff modelling to investigate the potential impacts of climate change on pluvial flooding in four European cities (Odense, Vienna, Strasbourg, and Nice). They showed increases in total area affected by SWF events, which varied between the cities due to a range of factors including differences in soils and topography. However, they only used change factors for extreme hourly precipitation derived from an ensemble of Global Climate Models (GCMs) downscaled via a 50-kmresolution RCM, which may not represent sub-daily extremes well.
Here, for the first time, a set of high-resolution (1.5 km and 12 km) RCM data has been used as input to a hydrological model to investigate the potential future changes in SWF over southern Britain. The aims of this paper are to • Investigate potential future changes in SWF hazard and impact, using surface runoff.
• Investigate the effect of RCM resolution on projections of change in surface runoff.
• Compare results for different parts of the country and in different seasons.
• Compare results based on surface runoff to those based on precipitation.
Section 2 describes the study area, models, datasets, and methods; Section 3 the results; and Sections 4 and 5 the discussion and conclusions.

| Study area
In England, SWF threatens more people and properties than any other form of flood risk; about 3 million properties are at risk of SWF, but about 2.7 million are at risk from rivers and the sea (Bevan, 2018).
Recent SWF events such as those in Hull (2007;Coulthard & Frostick, 2010), Newcastle (2012Newcastle City Council, 2013), andBirmingham (2016;Birmingham City Council, 2017)  has been developed that includes SWF as one of the hazards chosen for real-time pre-operational trials (Cole, Moore, Aldridge, Lane, & Laeger, 2013;Cole, Moore, Wells, et al., 2016;Speight et al., 2018). The SWF HIM system uses surface runoff estimates from the Grid-to-Grid (G2G) hydrological model (Bell, Kay, Jones, Moore, & Reynard, 2009;Moore, Cole, Bell, & Jones, 2006) to estimate the SWF hazard and links this to detailed inundation model outputs to assess impacts on property, people, transport, and infrastructure using a precomputed Impact Library (Aldridge, Gunawan, Moore, Cole, & Price, 2016). Here, a set of RCM data have been used to explore the potential future changes in SWF hazard, and property impacts, over the southern part of Britain ( Figure 1) at climate timescales.

| Regional climate models
In this study, a set of nested high-resolution Met Office Hadley Centre RCM runs are used. The 1.5-km model domain spans southern Britain ( Figure 1) and is driven by a 12-km RCM, which has a European domain and is in turn driven by a 60-km GCM (HadGEM3, Walters et al., 2011). The 1.5-km RCM is convection-permitting, meaning that it does not need to employ the convection parameterisation schemes required in coarser resolution models. The model runs cover both Current and Future periods (Table 1) and use climatological aerosols. Further RCM details are provided by Kendon, Roberts, Senior, and Roberts (2012) and Kendon et al. (2014).
Analyses using runs of the nested RCMs driven at the boundaries by ERA-Interim reanalysis data showed than the 1.5-km RCM better represents sub-daily precipitation in Britain than the 12-km RCM in terms of duration and spatial extent (Kendon et al., 2012) and summer extremes ; although, improvements in daily precipitation are less clear (Chan et al., 2013). Analyses using runs of the nested RCMs driven by GCM boundary conditions showed future increases in summer heavy rainfall in the 1.5-km RCM that were not seen in the 12-km RCM . Also, the 1.5-km and 12-km RCMs show different changes in summer precipitation extremes , with the 1.5-km RCM projecting increases of~10% in hourly intensity across a range of return periods (but little change in daily intensity), but the 12-km RCM projecting decreases in both hourly and daily intensity at short return periods (<5 years) and large increases at long return periods (>20 years).
Increases in (hourly and daily) winter precipitation intensity are much larger than for summer for both the 12-km and 1.5-km RCMs, and the increase in daily intensity is much higher for the 1.5-km RCM than the 12-km RCM. Kay, Rudd, Davies, Kendon, and Jones (2015) show that the 1.5-km RCM generally performs worse than the 12-km RCM for simulating river flows in 32 example catchments, with a clear east/west pattern of bias consistent with patterns of mean bias shown in the RCM precipitation data. The results using GCM-driven RCM runs show that in all seasons except summer the 1.5-km RCM tends towards larger increases in flood peaks than the 12-km RCM, with differences most pronounced for spring and winter.

| Hydrological model and driving data
The G2G is a distributed hydrological model that provides estimates of flow, surface runoff, and soil moisture on a 1-km 2 grid across Great Britain (Bell et al., 2009;Moore et al., 2006). An advantage of G2G is that it has one spatially consistent configuration and is able to model a wide variety of hydrological regimes due to use of spatial datasets (e.g. elevation, land cover, and soil type) in the model construction.
The effect of urban and suburban land cover on runoff and downstream flows is also included. The model addresses the ungauged hydrological forecasting problem and facilitates forecasting 'everywhere' (Cole & Moore, 2009) Price et al., 2012). Although the G2G has been used to assess the impact of climate change on river flooding (Bell et al., 2012;Bell, Kay, Davies, & Jones, 2016) it has not, until now, been used to analyse the impact on SWF.
G2G requires input time series of precipitation and potential evaporation (PE). Hourly precipitation is directly available from the RCM runs, but needs to be converted from the RCM grid (rotated lat-long) to the 1-km hydrological model grid (GB national grid). Conversion for the 1.5-km RCM uses area-weighting, whereas for the 12-km RCM, the data are copied to each of the corresponding 1-km grid boxes of the hydrological model grid. Hourly RCM precipitation is divided equally down to the 15-min model time-step.
Monthly PE is estimated from meteorological variables output by the RCMs, using the Penman-Monteith formula (Monteith, 1965). PE is divided equally down to the 15-min model time-step. A comparison of PE from the 1.5-km and 12-km RCMs shows they are very similar . Here, the PE from the 12-km RCM runs is also used for the equivalent 1.5-km RCM runs to ensure that any differences in surface runoff results are due only to differences in precipitation inputs. The estimation of PE for the Future period accounts for changes in stomatal resistance under higher atmospheric concentrations of CO 2 , which results in much smaller increases in PE than if fixed stomatal resistance values are applied . This has been shown to influence simulated future changes in river flows (Kay, Bell, Guillod, Jones, & Rudd, 2018) and could thus influence the simulated production of surface runoff.

| Analysis of precipitation and surface runoff
Data from the 12-km and 1.5-km Current and Future RCM runs (Table 1) are used to drive G2G. The 1-km grids of surface runoff simulated by G2G and the RCM precipitation are analysed relative to thresholds for 1-, 3-, and 6-hr duration extremes. Thresholds are defined as the 99.9th percentiles ( periods exceeding the threshold, which could count the same event multiple times).
As the 1.5-km and 12-km RCM runs show different precipitation changes in summer and winter , the analysis is also separated for the four seasons; winter (

| Analysis of property impacts
To analyse the potential impacts of SWF on property into the future, the analysis (Section 2.4) is repeated with spatially varying surface runoff thresholds from the Impact Library ( Figure 2) that utilise the updated Flood Map for Surface Water dataset (uFMfSW; EA, 2013).
The uFMfSW contains design flood maps from an inundation model for nine rainfall scenarios using combinations of three return periods (30, 100, and 1000 years) and three durations (1, 3, and 6 hr). The 'effective rainfall' used as input to the inundation modelling accounts for rural runoff processes and losses to urban drainage. The Impact Library provides the severity level (minimal, minor, significant, and severe) of property impact for each scenario, based on counts of properties at risk in each 1-km pixel (Table 2 of Aldridge et al., 2016).
Within this study, the three 1-hr impact maps are used to identify the minimum 1 hr 'effective rainfall' thresholds required to generate minimal, minor, significant, and severe impacts. If a pixel crosses the 'significant' severity level, it is implicit that it has already crossed the 'minimal' and 'minor' severity thresholds. Figure 2 shows that 'severe' and 'significant' property impacts can only occur in dense urban areas (e.g. London and Birmingham) and for relatively high threshold events, with lower threshold events causing lesser impacts ('minimal' or 'minor') in these areas. In more rural areas, higher threshold events can be needed to cause even 'minimal' or 'minor' impacts, and in some rural pixels there are too few/no properties to give any impacts. The G2G surface runoff estimates can reasonably be equated to the 'effective rainfall' estimates (Warren, Hunter, & Revilla-Romero, 2016), as done in the SWF HIM ; therefore, the effective rainfall grids can be used as spatially varying thresholds in the impact analysis. Due to the significant spatial differences in the thresholds, especially for higher severity levels, southern Britain is considered as a whole. The grids of exceedance counts are averaged over two regions, to the West and East of southern Britain (Figure 1). These regional averages and the percentage changes between the Current and Future time-slices, are plotted in Figure 4 for the 1-hr thresholds of precipitation and surface runoff ( The percentage changes in precipitation exceedance counts are all positive, higher in the Future time-slice than the Current time-slice, except for the 12-km simulation in the East in spring (~10% decrease).

| Precipitation and surface runoff
The largest percentage increases in precipitation exceedance counts are in winter and for the 1.5 km, but the exceedance counts in the Current period are very small in these cases. The smallest percentage increases in precipitation exceedance counts are in summer (1.5 km: 10% and 12 km:~50%). The percentage increases in precipitation exceedance counts in the West are higher than in the East, except in summer for the 12-km simulation. In winter and summer, the increases in precipitation exceedance counts in the 1.5-km simulation are larger than for the 12 km.
The percentage changes in surface runoff exceedance counts are smaller than for precipitation, in all seasons and in both the West and East. The percentage changes in surface runoff are positive, except for the 12 km in summer (West and East) and spring (East). The largest increases in surface runoff exceedance counts are in winter and for the 1.5-km simulation (~400% in the West and 500% in the East).
Decreases in surface runoff exceedance counts in summer are larger for the 12-km simulation than the 1.5 km. Percentage changes in surface runoff exceedance counts generally decrease slightly across the durations ( Figure S2). The uncertainty from individual years F I G U R E 2 Spatially varying, 1-hr, effective rainfall thresholds for four different severities of property impact (as estimated by jack-knifing) is relatively small (Figure 4a and Figure S1), especially for surface runoff. Figure 5 shows the 1-hr exceedance counts (totals for southern Britain averaged per year) and the percentage changes in those exceedance counts, for the four levels of property impact severity (Section 2.5). The exceedance counts for the 1.5-km simulation are higher than those for the 12-km simulation (Figure 5a), and summer counts show the greatest absolute difference between resolutions.

| Property impacts
The percentage changes in exceedance counts (Figure 5b) are positive and increase with increasing severity, except for the 12-km simulation in winter (not monotonic) and the 12-km simulation in spring where the percentage changes decrease with increasing severity and become negative for 'severe' impacts. In summer and autumn, across all severity levels, the 12-km percentage increases are higher than for the 1.5-km RCM, whereas in spring, the percentage increases for the 1.5-km RCM are higher than for the 12-km RCM (especially for the 'significant' and 'severe' impact levels). In winter, the percentage increases for the 1.5-km and 12-km RCMs are more similar to each other. The winter shows the largest percentage increases and also the lowest absolute exceedance counts, particularly for 'severe' impacts.

| DISCUSSION
The analysis in this study showed that the percentage changes in threshold exceedance counts for precipitation and surface runoff are Percentage increases in surface runoff are less than those of precipitation. This is likely due to the complex interaction of land cover (permeability) with surface runoff rates, and soil wetness will have an impact on how precipitation changes affect surface runoff changes.
This interaction with the ground is what will ultimately control whether there is surface water flooding, rather than the rainfall depth per se. Using the surface runoff approach outlined here therefore adds value over a purely rainfall-threshold-based method. G2G is a process-based model that uses spatial data to characterise heterogeneity in the runoff response and simulates greater runoff in urban areas, but differences between results for precipitation and surface runoff will depend on the way the hydrological model conceptualises/characterises runoff-production processes; results for other hydrological models may be different.
This study also showed that the largest percentage increases in both precipitation and surface runoff are in winter rather than summer, consistent with , suggesting a shifting seasonal balance of surface water flood risk. For property impacts, the largest percentage increases are also in winter; this follows-on from the largest increases in surface runoff.
However, for property impacts, it is the 12-km RCM that shows the largest percentage increase, especially for the 'minimal' and 'minor' severity levels. This may be related to scale as the 12-km surface runoff simulation may cover a wider area (i.e. more 1-km 2 impact cells) than the 1.5 km, which could affect the 'minimal' and 'minor' severity levels more due to there being more property impact cells at those levels than for 'significant' and 'severe' (Figure 2). Increased vulnerability for a particular season could affect the ability to respond and F I G U R E 4 (a) Regional average annual and seasonal exceedance counts of 1-hr precipitation (pr: squares) and surface runoff (sr: circles), for the 12 km and 1.

| CONCLUSIONS
This study produced the first estimates of future changes in surface water flood hazard and impact for southern Britain using a nationalscale gridded hydrological model and high-resolution RCM data. The approach analysed potential changes in the frequency of surface F I G U R E 5 (a) Annual and seasonal total (per year) exceedance of 1-hr surface runoff for the 12-km and 1.5-km RCMs (green and blue) for the Current (C) and Future (F) time-slices (filled and open) for four property impact severity thresholds. (b) Percentage changes in exceedance counts. Note that the bars go off the scale for severe impacts in winter. The totals are for southern Britain water flooding using percentile-based precipitation and surface runoff thresholds, as well as spatially varying surface runoff thresholds for property impacts. It was found that the largest percentage increases are for precipitation rather than surface runoff, in winter rather than in summer and projected by the 1.5-km RCM rather than the 12-km RCM. The largest percentage increases in property impacts are projected to be in winter. Future surface runoff estimates such as these could be used to supplement rainfall estimates used for sewer/drainage design (Dale et al., 2017).
It is important to recognise that surface water flooding can happen anywhere and is not always associated with urban areas and concrete. For example, there has been a dramatic rise in maize production to feed anaerobic digestion plants, and harvesting maize in late autumn (when the ground is often wet) can compact the soil meaning that rainwater is less easily absorbed and so more likely to produce surface water runoff and localised flooding (Bevan, 2018). Such changes would not be simulated by the hydrological modelling applied here, but changes in land-cover, including urbanisation could be accounted for in future work. Kaspersen et al. (2017) found that the relative influence of potential future climate change and recent historical urban development on pluvial flooding varied considerably between the four European cities they studied, so it would be interesting to investigate this balance in Britain. The inclusion of future urbanisation could be particularly important for flood damage assessments; Poelmans, Van Rompaey, Ntegeka, and Willems (2011) suggest that future fluvial flood risk could be influenced more by urban expansion than climate change for a small suburban catchment in Belgium.
The Impact Library applied here is based on a static set of receptor grids, but future work could allow for urban development.